The existing implementation will produce a cost of NaN if a tolerance of
infinity is entered, but the limit approaches zero. Being able to
specify that a state has no cost is useful, so this change adds support for
that.
It would crash in C++ if the global measurement was sooner than all the
snapshots.
Align Java with the changes and better document computation approach.
It was using the continuous B matrix to compute the feedback gain
instead of the discrete B matrix.
Tests were added for the matrix constructor overloads.
These classes are useful for storing previous robot positions to use in conjunction with the upcoming pose estimators.
Co-authored-by: Prateek Machiraju <prateek.machiraju@gmail.com>
Co-authored-by: Tyler Veness <calcmogul@gmail.com>
Co-authored-by: cttew <cttewari@gmail.com>
This also makes the Gradle build work with JDK 17.
The extra JVM args in gradle.properties works around a bug with spotless
and JDK 17: https://github.com/diffplug/spotless/issues/834
PMD.CloseResource was ignored because it's almost always a false
positive, and there are many of them.
The template argument order for UnscentedTransform was reversed to match
all the other UKF classes. Since UnscentedTransform is intended as a
class for internal use only, this shouldn't cause much breakage.
* Replace Matrix<> with Vector<> where vectors are explicitly intended.
I found these via `rg "Eigen::Matrix<double, \w+, 1>"`.
* Pass all Eigen matrices by const reference. I found these via `rg
"\(Eigen"` on main (the initializer list constructors make more false
positives).
* Replace MakeMatrix() and operator<< usage with initializer list
constructors. I found these via `rg MakeMatrix` and `rg "<<"`
respectively.
* Deprecate MakeMatrix()
This is an alternative to #2344 that handles arbitrary order derivatives
of arbitrary precision. The downside is that since it's part of
LinearFilter, it can't utilize the units type system in the same way to
make Calculate()'s input type different from its output type.
* Rename Butcher tableau sections in NumericalIntegration such that
top-left is c, top-right is A, and bottom-right is b
* Move edu.wpi.first.math.Discretization to
edu.wpi.first.math.system.Discretization
* Sort Java Discretization to match C++ function order
* Add tests for Java Discretization
* Required adding Runge-Kutta time-varying impl to tests
* Move C++ Runge-Kutta time-varying impl to tests only
* Users don't need it
Also refactored RKF45 implementation to match the new style, which is
easier to read.
The tests were switched from RKF45 to RKDP since it's more accurate.
Some valid warnings like throwing NullPointerException or using a for
loop instead of System.arraycopy() were fixed.
Abstract classes marked with PMD.AbstractClassWithoutAbstractMethod were
made concrete because they already had protected constructors.
Fixes#1697.
This also fixes a member function name inconsistency between languages
and adds missing documentation to C++'s KalmanFilterLatencyCompensator.
Fixes#3229.
The implementation of wpi::circular_buffer has been effectively replaced
with a dynamically sized copy of wpi::static_circular_buffer with a
resize() member function.
The units for angular Kv and Ka were inconsistent with the derivation. A
second factory function overload was added for angular units that uses a
trackwidth to convert to the other form.
Notice how section 15.2 of https://file.tavsys.net/control/controls-engineering-in-frc.pdf
defines the angular feedforward as u = Kv,angular v instead of u = Kv,angular + omega.
The units cancel for elements of A but not B, so just the B matrix was incorrect in our code.
This breaks existing C++ code since the units are part of the function
signature.
